diff --git a/benchmark/test/reference/distributed_solver.profile.stderr b/benchmark/test/reference/distributed_solver.profile.stderr
index 47918f049ea..907ff8a9c98 100644
--- a/benchmark/test/reference/distributed_solver.profile.stderr
+++ b/benchmark/test/reference/distributed_solver.profile.stderr
@@ -52,8 +52,34 @@ DEBUG: begin dense::fill
 DEBUG: end   dense::fill
 DEBUG: begin components::aos_to_soa
 DEBUG: end   components::aos_to_soa
-DEBUG: begin distributed_matrix::build_local_nonlocal
-DEBUG: end   distributed_matrix::build_local_nonlocal
+DEBUG: begin distributed_matrix::separate_local_nonlocal
+DEBUG: end   distributed_matrix::separate_local_nonlocal
+DEBUG: begin components::fill_array
+DEBUG: end   components::fill_array
+DEBUG: begin components::fill_array
+DEBUG: end   components::fill_array
+DEBUG: begin index_map::build_mapping
+DEBUG: end   index_map::build_mapping
+DEBUG: begin copy
+DEBUG: end   copy
+DEBUG: begin components::prefix_sum_nonnegative
+DEBUG: end   components::prefix_sum_nonnegative
+DEBUG: begin copy
+DEBUG: end   copy
+DEBUG: begin components::fill_array
+DEBUG: end   components::fill_array
+DEBUG: begin copy
+DEBUG: end   copy
+DEBUG: begin components::prefix_sum_nonnegative
+DEBUG: end   components::prefix_sum_nonnegative
+DEBUG: begin copy
+DEBUG: end   copy
+DEBUG: begin components::fill_array
+DEBUG: end   components::fill_array
+DEBUG: begin index_map::map_to_local
+DEBUG: end   index_map::map_to_local
+DEBUG: begin copy
+DEBUG: end   copy
 DEBUG: begin copy
 DEBUG: end   copy
 DEBUG: begin copy
diff --git a/benchmark/test/reference/spmv_distributed.profile.stderr b/benchmark/test/reference/spmv_distributed.profile.stderr
index d5e4913b51a..d3a645aa0f7 100644
--- a/benchmark/test/reference/spmv_distributed.profile.stderr
+++ b/benchmark/test/reference/spmv_distributed.profile.stderr
@@ -104,8 +104,34 @@ DEBUG: begin dense::fill
 DEBUG: end   dense::fill
 DEBUG: begin components::aos_to_soa
 DEBUG: end   components::aos_to_soa
-DEBUG: begin distributed_matrix::build_local_nonlocal
-DEBUG: end   distributed_matrix::build_local_nonlocal
+DEBUG: begin distributed_matrix::separate_local_nonlocal
+DEBUG: end   distributed_matrix::separate_local_nonlocal
+DEBUG: begin components::fill_array
+DEBUG: end   components::fill_array
+DEBUG: begin components::fill_array
+DEBUG: end   components::fill_array
+DEBUG: begin index_map::build_mapping
+DEBUG: end   index_map::build_mapping
+DEBUG: begin copy
+DEBUG: end   copy
+DEBUG: begin components::prefix_sum_nonnegative
+DEBUG: end   components::prefix_sum_nonnegative
+DEBUG: begin copy
+DEBUG: end   copy
+DEBUG: begin components::fill_array
+DEBUG: end   components::fill_array
+DEBUG: begin copy
+DEBUG: end   copy
+DEBUG: begin components::prefix_sum_nonnegative
+DEBUG: end   components::prefix_sum_nonnegative
+DEBUG: begin copy
+DEBUG: end   copy
+DEBUG: begin components::fill_array
+DEBUG: end   components::fill_array
+DEBUG: begin index_map::map_to_local
+DEBUG: end   index_map::map_to_local
+DEBUG: begin copy
+DEBUG: end   copy
 DEBUG: begin copy
 DEBUG: end   copy
 DEBUG: begin copy
diff --git a/benchmark/test/reference/spmv_distributed.profile.stdout b/benchmark/test/reference/spmv_distributed.profile.stdout
index 5eb5710a134..bbef87d0b89 100644
--- a/benchmark/test/reference/spmv_distributed.profile.stdout
+++ b/benchmark/test/reference/spmv_distributed.profile.stdout
@@ -5,7 +5,7 @@
         "comm_pattern": "stencil",
         "spmv": {
             "csr-csr": {
-                "storage": 6564,
+                "storage": 6420,
                 "time": 1.0,
                 "repetitions": 1,
                 "completed": true
diff --git a/benchmark/test/reference/spmv_distributed.simple.stdout b/benchmark/test/reference/spmv_distributed.simple.stdout
index 1bb9378eb7c..77bdef168d3 100644
--- a/benchmark/test/reference/spmv_distributed.simple.stdout
+++ b/benchmark/test/reference/spmv_distributed.simple.stdout
@@ -5,7 +5,7 @@
         "comm_pattern": "stencil",
         "spmv": {
             "csr-csr": {
-                "storage": 6564,
+                "storage": 6420,
                 "max_relative_norm2": 1.0,
                 "time": 1.0,
                 "repetitions": 10,